Individual and Province Inequalities in Health among Older People in China: Evidence and Policy Implications

This paper uses multi-level modelling to analyse data from the nationally-representative Chinese Health and Retirement Longitudinal Study (CHARLS) in order to investigate the characteristics associated with poor health among older people, including individual and household characteristics as well as the characteristics of the provinces in which the older person lives (contextual effects). The results show that older Chinese women, rural residents, those with an education level lower than high school, without individual income sources, who are ex-smokers, and those from poor economic status households are more likely to report disability and poor self-rated health. Differentials in the health outcomes remain substantial between provinces even after controlling for a number of individual and household characteristics. Improvements in life expectancy in China over the past 50 years combined with marked declines in fertility have resulted in rapid population ageing, reflected in an increase in the absolute and relative number of older people in the population and resulting in a challenge for the design of adequate social policies in health and social care (Cai et al., 2012; Woo et al., 2002). Moreover, these demographic changes have taken place alongside significant social and economic developments which are further reshaping China's population. Before 1976, which marked the death of Chairman Mao, " economics gave way to politics " and the state had a strong mandate to run the economy and organise production for the benefit of society (Chen, 2002: 571). During this time, and typically for a centrally-planned economy, Chinese citizens were promised guaranteed employment (often referred to in popular culture as the " iron rice bowl "), egalitarian distribution of resources and outputs, and cradle-to-grave welfare coverage (Li, 2012). The transition from a socialist to a market economy, which accelerated from the late 1970s onwards, created opportunities for the development of social welfare services and the solution of fundamental social problems; however inequalities between occupational groups and regions emerged as " new " social risks facing the government. Recent market reforms, decentralisation and economic globalisation have impacted different social groups, regions and industries unevenly (Zhu, 2013) with the resultant rural–urban migration rapidly altering the demographic composition of different regions in China. Increasing rural–urban migration has served to further emphasise the dual policy challenge of health and social provision for low-paid migrants in cities, and for older people " left behind " in rural areas often caring for grandchildren (Biao, 2006). Against the background of …


The effect of individual characteristics on health status
A large body of existing literature from Europe and North America has evidenced the association between a range of demographic and socio-economic characteristics and an individual's health status; however such research evidence is still scarce in the Chinese context. The Marmot Review (2010) in the British context revisited the strong link between socio-economic status and the reporting of poor health status, while evidence from other countries of the developed world is compatible (Hay, 1988;House et al, 1990). Although the direction of the causal mechanism between poor socio-economic status and poor health status is the subject of on-going research and debate, the evidence of the association between the two concepts, regardless of the way they are operationalised in empirical research (eg. socioeconomic status as individual income, and health status as self-reported health), is not disputed. The evidence in the Chinese context is still relatively scarce, and also presenting a more complex picture as a result of the particular socio-political factors which have shaped demographic patterns, the provision of healthcare services and access to such services alike.
For instance, Liang et al (2000) examined this relationship in the Wuhan province and found a socio-economic gradient in the report of poor health status, with individuals in lower socioeconomic classes being more likely to report poor health than those in higher classes. More recently, Zimmer and Kwong (2004) found that more 'traditional' socio-economic indicators such as income and education, were relatively weak predictors of the report of poor health status, while banks savings and pension eligibility indicateded a stronger effect. Finally, Lowry and Xie (2009) argue that although socio-economic status is positively and strongly associated with health status for individuals in younger ages, such an association is weaker in the latter part of the life course.

The effect of province characteristics on individuals' health status
Contextual and compositional effects on individuals' health status have been evidenced by previous research (e.g. neighbourhood effects (Becares et al, 2012), deprivation (Jones et al, 2000), income inequality (Feng et al, 2012)). The interaction of such factors is important to examine as China is highly spatially differentiated in terms of its economic development and social security, resulting in variability in the quality and availability of health facilities between provinces. Previous studies have found that income inequality is strongly associated with the reporting of poor self-rated health among elderly persons (Feng et al, 2012), and that there was no evidence of a significant improvement in the health of elderly persons living in provinces with better health facilities (Feng et al, 2013). In an earlier study, Yin and Lu (2007) found that the prevalence of medical conditions at the province level had an impact on elderly persons' report of disability, defined as difficulty with specific Activities of Daily Living (ADLs).
This paper aims to contribute to the literature by investigating the health outcomes of older people in China and examining the extent to which these outcomes are influenced by individual and province level characteristics. The paper addresses the following central research question: how does the health of older people vary according to demographic characteristics, socio-economic indicators, health risk behaviours, household/family factors and provincial level factors? The next section discusses the data and methods to be used. The results of a series of multivariate regression models are then presented, followed by a discussion of the results, drawing out the implications for policy makers.

Data
The analysis in this paper combines data from two different sources to explore the impact on the health of older people of individual, family/household and province level characteristics.
The province information used in the analysis of this paper comes from the Chinese statistics yearbook in 2012. Nineteen variables were chosen to reflect the contextual effects of each province including urbanisation, economic development, marketisation, spending on health care, health facilities and quality of living (Table 1). (Table 1 about here) The individual and family level data are from the wave 1 of the national baseline of the China Health and Retirement Longitudinal Study (CHARLS) conducted in 2011-2012. The CHARLS survey covers 450 villages/urban communities in 150 counties/districts, located in 28 provinces across the country 1 . It is based on a randomly selected sample of people aged 45 and over in the household. Having identified households that include an eligible member (age 45 and over), he/she was defined as the main respondent. Where the household had more than one age-eligible member, the respondent was randomly selected. The data used in this study includes 10,717 interviewees aged 50 and above with complete responses. The dataset has a natural hierarchal structure with individuals nested within families/households within provinces.
Five health indicators are considered: disability (difficulty performing Activities of Daily Living (ADLs) or Instrumental Activities of Daily Living (IADLs)), subjective health status (self-reported health (SRH)), perceived life satisfaction and self-reported memory) 2 . The indicator of difficulty with ADLs includes the six basic activities of dressing, bathing, eating, getting into/ out of bed, using the toilet, and controlling urination and defecation. The response categories in this indicator were: No difficulty at all; Have difficulty but can still do it; Have difficulty and need help; and Cannot do it even with help. Approximately 82 per cent of the respondents reported no difficulty at all with any of the six ADL. Therefore, a binary variable was constructed after counting the number of ADLs a respondent reported difficulty with, with zero representing no difficulty at all for any of the six ADLs, and one representing any difficulty with any of the six ADLs. A similar process was followed for the indicator of difficulty with any of the five IADLs 3 . Self-reported health is a subjective report of one's health, and is reported on the following scale: very good, good, fair, poor, or very poor. A derived variable was constructed with 'positive' categories in the first category (very good or good), fair in the second category, and 'negative' categories in the third (poor and very poor).
This process was also used for the construction of derived variables for the indicators of life satisfaction and self-reported memory. The distribution for these health indicators across the sample is shown in Table 2.
The predictor variables are categorised into three groups: firstly, individual demographic, socio-economic, social security and health behaviour characteristics (age, gender, marital status, urban/rural residence, education, income sources, has/has not medical insurance and smoking status); secondly, household/family characteristics (self-rated living standards, and 2 A sixth outcome indicator was tested in the analysis (report of difficulty with mobility functions), however this model did not produce significant results (see Appendix 1 and 2). 3 There are five questions in the CHARLS dataset relating to experiencing difficulty with IADLs, doing household chores; preparing hot meals and shopping; shopping for groceries; managing your money (e.g. paying your bills, keeping track of expenses, or managing assets); and taking medications. whether the household receives Dibao 4 ); and thirdly, the province-level variables. The distribution of individual variables across the sample is shown in Table 2.

Methods
Nineteen variables are chosen to reflect the province characteristics, and some of these variables are highly correlated (e.g. the correlation between the percentage of urban population in a province and GDP per capita is 0.935). This indicates significant multicollinearity in the model. Thus, in order to capture overall province characteristics, we used factor analysis to generate a summary factor score for each province (Johnston, 1978). The results of this factor analysis are discussed in the Results section below.
Since the dataset has a natural hierarchal structure with individuals nested within provinces and the aim of this paper is to analyse the effects of individual characteristics and province characteristics on the health indicators simultaneously, multilevel logistic regression models are appropriate (Hox, 2002). In terms of the different number of categories in the health indicators, binomial logistic regression is used to examine the determinants of reporting difficulty with ADLs and IADLs, while multinomial logistic regression is used to examine the determinants of self-reported health, life satisfaction and self-reported memory. All the models were estimated using the MLwiN 2.27 software (Rasbash et al., 2009) estimation is used as it can decrease the inherent bias associated with using maximumlikelihood procedures for binary/multinomial models (Browne and Draper, 2006).

Province characteristics
The factor analysis of the 19 variables at province level generated three factors based on the number of the Eigenvalue that exceeds 1.0, with a communality of 78.8 per cent, which means that the majority of observed variances of the data could be explained by these three factors. Sorted rotated factor loadings and communalities for the variables are shown in Table   3. Factor loadings of less than 0.6 (only 36 per cent of variance in common) were set to 0.
( Table 3 about here) From Table 3, it can be seen that 52 per cent of the observed variability of the original variables is accounted for by factor 1. This factor was labelled as a province being 'Developed, marketised and lower level of spending on health care and lower provision of health facilities (D&M)', since it refers to provinces with a higher level of urbanisation, GDP per capita, migrant rates, average income, VAT per capita and foreign investment, a higher proportion of private or foreign and overseas Chinese industry employees, but a lower proportion of state-own employees, lower expenditure for medical and health care in total revenue of province and a low level of health care institutions per 10,000 population.
Approximately 16 per cent of the observed variance is accounted for by factor 2. This factor was labelled as 'Higher level of health facilities and quality of life (HLQ)', since it represents a higher number of health care facilities and lower Engel's coefficients in urban and rural areas. Finally, approximately 10 per cent of the observed variance was accounted for factor 3, which was labelled as 'Strong state influence and social security (SSI)', reflecting provinces which have a higher proportion of state owned enterprises in the total of fixed asset investment and higher percentage of expenditure for social safety net. Table 4 presents the factor scores for the provinces. From the labelled factors, it is possible to categorise the province characteristics. For the first factor (D&M), Shanghai is the most developed and marketised with lower health care and facilities province, compared to Hebei which is the least developed one. In terms of the second factor (HLQ), Beijing shows the best quality of life and better health facilities, while Yunnan shows the worst one. In terms of Strong State Influence, Qinghai has the strongest state influence and social security, while Shandong is the province with the weakest state influence and social security. In order to have a clearer visualisation of the distribution across these factors, Figures 1A to 1C map the three factor scores for each province with darker colours representing a higher factor score.  In order to have an understanding of the distribution of older people across China, Figure 1 also shows the percentage of 65+ year old people ( Figure 1D) in each province, indicating that Shandong, Jiangsu, Liaoning and Sichuan are the 'oldest' provinces in the country. This shows also that the highest concentration of older people in China tends to be in central and eastern provinces, where provinces report a high degree of marketization and economic development, coupled with a relatively low level of healthcare services, both of which can impact on the health status of older people.
( Figure 1 about here) Table 5 brief summarizes the significant effects of predictors on the health outcomes. In this study, there is only a linear relationship between age and some health outcomes (no quadric effects are found). Educational qualification and smoking status have significant effects on all health outcomes, while only two significant effects of whether having medical insurance are found on IADLs and self-reported health. There are some significant effects of province characteristics on ADLs, Life Satisfaction and Self-reported memory.

Multilevel analysis results
( Table 5 about here) Tables 6-7 present the results for the binary multilevel logistic regression models of reporting difficulty with ADLs and IADLs, while Tables 8-10 present the results for the multinomial multilevel logistic regression models of self-reported health, life satisfaction and self-reported memory. All the results are shown as odds ratios. In order to compare between the effects of individual and household characteristics and province characteristics on health outcomes, the tables present two sets of results: one shows the individual and household effects, and the other shows the additional contribution of province effects.
In terms of the effect of individual characteristics on one's difficulty with ADLs, Table 6 highlights that individual demographic characteristics have an effect, with the risk of reporting such difficulty increasing by age, as every year of age increases the odds of reporting a difficulty with an ADL by 0.05. Females are 35% more likely to report a difficulty with ADLs than males, however there are no substantial differences according to marital status. Variables reflecting one's socio-economic status are an important part of the determinants of reporting difficulty with ADLs. Individuals living in rural areas are more likely to report difficulty with ADLs than urban residents (ORs=1.44), while those with higher educational qualifications (high school and above) are associated with a lower risk of reporting difficulty with ADLs. Income sources are also important in explaining difficulty with ADLs, as individuals receiving income from wages are the least likely to face a risk of difficulty with ADLs compared to individuals receiving income from other sources, while the receipt of Dibao by the household was strongly associated with the report of difficulty with ADLs. No substantial differences are found between individuals having medical insurance or not.
The subjective economic status of the family also contributes to one's individual risk of reporting difficulty with ADLs, as those who rated their standard of living as low showed higher odds than those reporting a high standard of living of reporting difficulty with ADLs (ORs=1.71). Finally, health-risk behaviour was also part of the explanation, as ex-smokers show higher odds of reporting difficulty with ADLs than those never smoke (ORs=1.45), and at the 95% level, ex-smokers are also significantly more likely to report difficulty with ADLs than those who are current smoking. This result may indicate a more complex effect of differential level or amount of smoking among ex-smokers, which negatively impacts on their chances of reporting difficulty with ADLs.
No substantial differences were found in the individual effects on the risk of reporting difficulty with ADLs when the province characteristics were added to the model. Only the D&M has a significant effect on an individual's risk of reporting difficulty with ADLs, with persons living in provinces scoring higher in terms of developed, marketised, low health expenditure and health care institutions province being less likely to report difficulty with ADLs. Table 7 shows the individual and province effects on one's risk of reporting difficulty with IADLs. No substantial differences are found among demographic factors (e.g. age, gender, marital status and urban/rural resident). There is a slight effect of socio-economic characteristics on the risk of reporting difficulty with IADLs, in that individuals who have no educational qualifications are less likely to report difficulty with IADLs compared to those with qualifications at the high school level and above (ORs=0.86). No substantial differences are found among individuals receiving income from different sources, while individuals without medical insurance are more likely to report difficulty with IADLs than those who have medical insurance (ORs=1.37). Ex-smokers show lower odds of reporting difficulty with IADLs compared to those who have never smoked (ORs=0.84). Finally, no substantial differences are found according to the level of living standards perceived among individuals or whether the household receives the Dibao benefit. Interestingly, no province effects are found in the risk of reporting difficulty with IADLs. Table 8 presents the multinomial multilevel regression of reporting positive, fair or negative self-rated health. There is significant positive effect of age on reporting negative SRH (ORs=1.02), while females are more likely to report fair and negative SRH than males. In terms of marital status, widowed persons are less likely to report negative SRH than married persons, indicating the beneficial effects of cohabitation. Those living in urban areas, who have higher educational qualifications, who receive income from wages, and who do not have medical insurance are more likely to report positive SRH. Ex-smokers show 71% higher odds of reporting negative SRH than those never smoke, while individuals reporting a low standard of living and the receipt of Dibao by their household are more likely to report fair or negative SRH. Finally, no province effects are found in the risk of reporting fair and negative SRH. Table 9 demonstrates multinomial regression models for reporting perceived wellbeing. Age shows a negative association with reporting fair or negative wellbeing, however no substantial differences are found between women and men in this respect. Widowed persons are less likely to report fair wellbeing than those who are married, while those in 'other' categories of marital status show 83% higher odds of reporting negative life satisfaction than married persons. Individuals living in rural areas are 19% more likely to report negative life satisfaction than those living in urban areas. Socio-economic status shows a significant effect on life satisfaction, as illiterate individuals or those with educational qualifications below the primary level are less likely to report fair life satisfaction than those whose qualifications are at the high school level or above; however, no significant difference is found between Finally, Table 10 illustrates the multinomial regression for reporting positive, fair or negative self-reported memory. Individuals who are older, female, married, living in rural areas, with lower educational qualifications, no income sources or receiving income from 'other' sources, and who are ex-smokers, are more likely to report fair/negative memory than those who are younger, male, widowed, living in urban areas, with an education at the high school or above, receiving income from wages and never smoke. The lower perceived living standard again shows a positive association with individuals' report of fair or negative memory. Province characteristic appears to have an effect on an individual's risk of reporting fair memory, those living in provinces with good health facilities and quality of life being less likely to report fair memory (ORs=0.79).  (Figure 6).

Discussion and conclusion
The aim of this paper was to contribute to our understanding of how the combination of individual and province characteristics can affect an individual's risk of reporting poor health status according to a range of health indicators. Overall the results show that specific demographic and socio-economic characteristics are strongly associated with the report of poor health status. More specifically, the analysis in this paper show that being female, living in rural areas, having low educational qualifications, having no income sources or receiving income from 'other' sources, being an ex-smoker, reporting relatively low living standards and living in a household which reports the receipt of Dibao, are characteristics associated with the report of difficulty with ADLs or IADLs, fair/negative SRH, fair/negative life satisfaction and fair/negative memory. However, such individual characteristics are not the whole story, and this paper has sought to understand the effect of the province in which older persons live, on their risk of reporting poor health status. In this respect, the results show that individuals living in provinces with better health facilities and better life quality are less likely to report difficulty with ADLs or negative life satisfaction.
The results point to significant policy implications which relate both to improving the immediate environment in which the individual's lives, such as their family and household, and to the wider environment or province where their household is located. Lower socioeconomic status, whether measured through individual or household indicators, is clearly and strongly associated with the report of poorer health status, as is health-risk behaviour measured in this study through smoking status, which itself has been independently associated with both lower socio-economic status and with poorer health status (Jarvis and Wardle, 1999). Public health campaigns against smoking, which are aimed at both younger and older cohorts of individuals, can contribute to the improvement of health status of current and future generations of Chinese individuals. However, statutory assistance in the form of both cash benefits and health services for individuals at the lowest part of the income distribution may also add to a comprehensive set of social policies protecting the most vulnerable groups in society. Finally, differences between provinces in a range of indicators, such as modernisation and the amount of expenditure on health services, reflect a need for greater redistribution of resources on the part of the state from those provinces with higher resources, to those with lower resources. Similarly, such an adjustment also requires ensuring that rural parts of provinces are not disadvantaged compared to urban parts. This is particularly important as economic migration has resulted in a higher concentration of older individuals in rural parts of mainland China. *Migrants refer to those who have lived in a place where their household registration does not belong. ** Engel's coefficient is a measure of the percentage of food expenditure in the total of personal consumption expenditure. This is between 0 and 100, and the closer to the latter, the poorer an individual is.    Notes: " " denotes significant effects; "-" denotes non-significant effects. * There is no substantially different effect of individual and province characteristics on the difficulty with mobility; therefore, the results are shown in the Appendix 1 & 2.